Development of Two Robust Data-Driven Models Using Machine Learning and Artificial Neural Network Methods to Predict Wells Fluid Rate in a Challenging Offshore Brown Field

Abstract

Abstract Two robust data driven models based on machine learning (ML) and artificial neural network (ANN) methods were introduced to overcome the shortcomings of physical and virtual well testing in Gulf of Suez offshore fields. The aim of these new models is to use the existing data and create a precise/easily accessible tool that fill the gap in well monitoring and testing system to predict wells fluid rate, improve field optimization and properly allocate oil production. A comprehensive methodology was applied to build/verify a robust virtual model as following: 1) Analyzing General Energy Equation to select the relevant inputs, 2) Data Collection, 3) Exploratory Data Analysis (EDA), 4) Feature Engineering, 5) Machine Learning Model Selection, 6) Hyper-parameters Fine Tuning, 7) Developing Artificial Neural Network model, 8) Models Deployment. Exploratory Data Analysis (EDA) and the General Energy Equation were used to select ten main parameters affecting the model’s accuracy. The selected features include wellhead pressure, wellhead temperature, reservoir temperature, reservoir pressure, water gravity, difference between reservoir and bubble point pressure, watercut percent, injection gas, downstream pressure, and tubing type. Different machine learning models based on linear, support vector machine, decision trees and gradient boosting methods were programmed. The results of these models were compared based on coefficient of determination (R2 score), root mean square error (rmse), mean absolute error (mae), and mean absolute percentage error (mape). XGboost regressor was selected as the best model, then the model hyper parameters were fine-tuned using grid search method. The final model results of test dataset showed R2 score, rmse, mae and mape of 0.9674, 323, 227 and 13.1% respectively. Furthermore, ANN was created and fine-tuned to select the model architecture. The model was evaluated using the same train and test data where the model showed comparable results to the best ML models. The results of ANN model showed R2 score, rmse, mae and mape of 0.9603, 357, 241 and 13.7%. respectively.